Supermicro rolls out H15 portfolio on AMD EPYC 9006
Supermicro announced its next-generation H15 server portfolio powered by 6th Gen AMD EPYC 9006 Series CPUs, targeting cloud, enterprise, storage, HPC, and agentic AI workloads. Systems support up to 256 cores and 512 threads, with the vendor claiming a 1.7x generational CPU performance gain and expanded memory and I/O bandwidth. The line-up spans Hyper dual-socket enterprise nodes, CloudDC systems built on the OCP DC-MHS specification, GrandTwin 2U four-node platforms, FlexTwin 1OU two-node liquid-cooled nodes, Petascale Storage systems reaching 4.8PB per chassis, and the H15 8U SuperBlade.
Complementing the CPU portfolio, Supermicro expanded its AMD GPU-powered AI infrastructure with 5U PCIe GPU servers supporting up to ten AMD Instinct MI350P GPUs per chassis, and the 72-GPU AMD Helios Platform shown at Computex 2026 for large-scale AI training. The combined release signals Supermicro's continued bet on AMD's CPU-plus-GPU stack as competition with NVIDIA-centric designs intensifies.
ASRock Rack crams eight Blackwell Ultra GPUs into a 4U chassis
ASRock Rack launched the 4U16X-GNR2, a 4U server pairing two Intel Xeon 6 host processors with eight NVIDIA Blackwell Ultra GPUs linked via onboard NVLink. The chassis adds 12 U.2 NVMe bays, eight 800 Gbps OSFP network cages for 6.4 Tbps of front-panel fabric, and ten 3 kW 80 Plus Titanium PSUs delivering 30 kW of installed nameplate capacity. ASRock Rack offers a direct liquid-cooling variant and a ZutaCore two-phase model for facilities avoiding water loops, reflecting growing operator sensitivity to data center cooling constraints.
The system targets AI training and inference clusters that want a familiar eight-GPU topology per node, local NVMe storage, and dense packaging, though the 30 kW power and liquid-cooling requirements will shape rack planning and facility design.
Meta and BlackRock commit $14bn to an El Paso AI campus
Meta folded its half-built El Paso data center into a joint venture with BlackRock, with BlackRock-managed funds taking 80% and Meta retaining 20%. Total development is valued at roughly $14bn, split between BlackRock's $4.9bn cash contribution at financial close, Meta's land and construction-in-progress valued at about $2.3bn plus a $1bn distribution, and roughly $12.5bn of project-level debt that does not sit on Meta's balance sheet. Meta remains construction manager, property manager, and sole tenant once servers come online in 2028, with an initial four-year lease and four extension options.
The structure mirrors Meta's earlier 80/20 sale-leaseback for its Hyperion campus in Louisiana and lets Meta book the site as an operating lease rather than capital expenditure, a distinction relevant to its $125bn–$145bn 2026 capex guidance. The El Paso campus is designed to deliver one gigawatt of compute, though the long-dated debt against short-depreciating server hardware remains a financing risk the partners have not addressed publicly.
Intel's datacenter CPU business rebounds on agentic AI demand
Intel said its Q2 2026 datacenter results showed the highest growth in server CPUs in fifteen years, attributing the rebound to demand for AI host CPUs and for CPU clusters running the Python sandboxes that execute agentic AI workflows. The commentary positions general-purpose CPUs, rather than accelerators, as the unexpected beneficiary of the GenAI cycle, even as AMD's EPYC line remains performance-competitive and hyperscalers expand homegrown Arm designs.
CEO Lip-Bu Tan told Wall Street that Intel remains on track for 14A risk production for internal products in the second half of 2027, with high-volume ramp in 2028, and cited increasing customer momentum around the 14A node. Tan also highlighted interest in Intel's EMIB-T packaging as TSMC's CoWoS-L capacity stays constrained, while CFO David Zinsner pointed to growing custom CPU and IPU engagements as an adjacent revenue line.
Share this article
